Skip to content

Olala mini: Europe's largest cloud provider unveils its first AI model

Comic-style engineer plugs a cable into a small glowing AI brain box in a vaulted server hall, a search lens sparking above Models

Olala mini is the first AI model from OVHcloud, Europe's largest cloud provider. The French company announced it on November 17. The open model has 7 billion parameters, uses only 1 billion per request and runs on a GPU with 8 GB of memory. OVH trained it from scratch on three European public supercomputers.

The weights are not out yet. OVH says it will publish them on Hugging Face under the Apache 2.0 license. The official launch is set for its OVHcloud AI Summit in Paris on November 19, and French media expect the weights that day. Some early reports said the model was already released, but for now only the announcement is public.

What OVH built with Olala mini 7A1B

The full name, Olala mini 7A1B, describes the design: 7 billion parameters in total, about 1 billion active at a time. The model is a mixture of experts with 256 experts per layer. Only a small part of the network works on each request, which makes it faster and lighter on memory.

OVH calls the architecture new. It combines two ideas:

  • Differential-TPA, a hybrid attention method that cuts the memory needed during inference and helps the model focus on relevant information;
  • Mamba-3 layers, which keep a fixed-size summary of the conversation instead of growing memory with every message.

OVH has not explained yet how the two parts fit together. The technical report should arrive with the weights.

The design pays off in speed, according to OVH. The model reaches 160 tokens per second on an NVIDIA RTX A4500. On an ordinary CPU with no GPU at all, it still manages about 50 tokens per second. Its context window is 32,000 tokens.

How OVH trained Olala on European supercomputers

OVH trained the model on three public supercomputers: Jupiter in Germany, MeluXina in Luxembourg and Jean Zay in France. They use NVIDIA GH200, A100 and H100 chips. The public support equals €7 million in computing hours, which OVH calls decisive.

The project also created a new in-house team, OVHai LLM. OVH promises to release more than the weights: the pre-trained base model, a technical report and documentation for compliance with the EU AI Act. It has not said whether the training code or datasets will be published.

OVH CEO Octave Klaba explained why a cloud company now trains its own models. "We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come," he said. He added that training costs have fallen twentyfold in three years, so building models can now be profitable.

The first customer is already in place.

Qwant already uses Olala mini

Qwant, the French search engine, uses Olala mini for its flash answers, the short AI replies in search results. OVH says answers became 20% faster and running costs fell to a third, compared with the model Qwant used before. Qwant previously relied on models from Mistral, French tech site Next reports.

There is a family link. Since 2023, Qwant has belonged to a company co-owned by the Klaba family and Caisse des Dépôts, a French public financial institution, Next reports.

What comes next for Olala

OVH laid out a roadmap:

  • Olala 1.1, expected by the end of 2026, will extend the context window to 128,000 tokens and improve math reasoning.
  • Olala 2.0 will exceed 100 billion parameters and split into specialized models for business agents, cloud DevOps and cybersecurity, both defensive and offensive.

The agent version should power OVHai Workspace, OVH's planned alternative to Microsoft 365 and Google Workspace. Its encrypted version will need models that run locally on the user's computer. That explains why the first Olala is so small.

Is Olala mini a real rival to Mistral

Not yet, and it is not trying to be one. OVH published speeds but no benchmark scores, so its quality is unknown until outside tests arrive. With 1 billion active parameters and a 32,000-token window, Olala mini is built for fast, cheap tasks such as short answers and data processing. Mistral's flagship models, such as Mistral Large 4, play in a different league.

Its importance lies elsewhere. A European cloud giant has now trained its own model from scratch with public computing power and plans to release it under a permissive license. If Olala 2.0 reaches 100 billion parameters as planned, Europe will have a second serious open model maker next to Mistral. The real test starts once developers can download the weights and check OVH's claims.

Sources

Portrait of Felix Laurent, author at thehype.news

Felix Laurent

News Writer at thehype. Felix is a journalist with experience in marketing. He covers artificial intelligence and emerging technologies, focusing on the ideas, innovations, and trends shaping the future of tech.

More articles by Felix Laurent →
On air · radio.thehype.news ↗ AI news radio — 24/7